fix(ci): stop flaky sleep test and compat evaluation errors

[jariy17]

Fixes two CI failures seen on #674 and #675. Neither was caused by those PRs.

1. Flaky test_zero_delay_skips_sleep

Fails with Expected 'sleep' to not have been called. Called 898 times.

Why:

• Importing ragas starts a background thread that calls time.sleep(1) in a loop. RAGAS_DO_NOT_TRACK doesn't stop it.
• Our tests patch paths like middleware.time.sleep. That patches the shared time module, so every thread sees the mock.
• If the ragas thread wakes while a test holds the patch, the mock returns instantly, the loop spins, and the mock records hundreds of calls.

Fix: tests/conftest.py stops the ragas thread after the first test that imports ragas.

2. Compat (evaluation) errors

test_third_party_adapters.py needs OpenAI and Braintrust API keys. The compat job doesn't have them, so its 16 tests always error. This PR skips them in that job; the secure integ workflow still runs them.

Follow-up: add skip guards to the test file itself so this ignore can be removed.

Testing

• Unit tests: 3565 passed. 4 memorystore typing tests fail locally with or without this change.
• The ragas thread is gone at the end of the run with the fix, and still running without it.
• Compat evaluation collects 37 tests instead of 53. The 16 dropped are the ones that errored.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)